融合大语言模型与提示学习的土壤分类知识图谱构建

Construction of soil classification knowledge graph integrating LLM and prompt learning

关键词:  土壤分类  知识图谱  大语言模型  智能问答  提示学习

Key words:  soil classification  knowledge graph  LLM  intelligent QA  prompt learning

摘要:

【目的】 针对土壤数据异构分散、分类知识界定模糊、层级逻辑脱节以及通用大模型本土化语义理解不足等问题，构建专业土壤分类知识图谱，并开发SoilGraphRAG系统，实现知识的深度关联与精准推理。【方法】 文章采用“文档智能解析+专家规则注入”的双重策略，结合MinerU解析与滑动窗口策略，从标准文档中自动化抽取“实体—关系—属性”三元组，构建包含6级分类体系的土壤知识图谱并存入Neo4j。并基于LangGraph设计ReAct智能体，构建“语义向量+Cypher结构化”双通道检索机制。【结果】 （1）成功构建了覆盖全量土壤分类单元的高质量图谱。基座模型QwenLong-L1-32B的检索召回率达89.65%，复杂推理能力得分为87.5，均优于DeepSeek-R1-Distill-Qwen-7B等对比模型。（2）得益于图谱结构化约束，系统的幻觉抑制能力评分高达86.8，验证了该框架在保障知识精确度方面的优势。（3）“向量语义检索+Cypher结构化查询”的双通道机制在不同题型下表现互补，结合ReAct智能体调度后显著提升了模糊语义与复杂逻辑问题的综合响应能力。【结论】 该研究提出的框架成功实现了土壤分类标准知识的结构化重组与智能问答，为土壤分类体系的深度挖掘提供了路径，也为其他垂直领域知识图谱的构建提供了新范式。

Abstract:

[Purpose] To address issues of heterogeneous and fragmented soil data，ambiguous classification knowledge boundaries，disconnected hierarchical logic，and insufficient localized semantic understanding of general-purpose large language models，this study develops a domain-specific soil classification knowledge graph and the SoilGraphRAG system to enable deep knowledge association and precise reasoning.[Method] A dual strategy combining intelligent document parsing and expert rule injection was adopted. By integrating MinerU parsing with a sliding window approach，entity-relation-attribute triples were automatically extracted from standard documents to construct a six-level soil classification knowledge graph stored in Neo4j. A LangGraph-based ReAct agent was designed to support a dual-channel retrieval framework integrating vector-based semantic retrieval with Cypher-based structured querying.[Result] （1）A high-quality knowledge graph covering the full set of soil classification units was successfully constructed. Using QwenLong-L1-32B as the backbone model，the retrieval recall reached 89.65% and the complex reasoning score was 87.5，both outperforming baseline models such as DeepSeek-R1-Distill-Qwen-7B. （2）Benefiting from the structural constraints of the knowledge graph，the system achieved a hallucination-resistance score of 86.8，demonstrating its advantage in ensuring knowledge accuracy. （3）The dual-channel mechanism combining vector-based semantic retrieval and Cypher-based structured querying exhibited complementary strengths across different question types；when coupled with ReAct agent scheduling，it significantly improved the overall response capability for ambiguous semantic queries and complex logical reasoning tasks.[Conclusion] This framework effectively organizes soil classification standards and supports precise semantic retrieval and reasoning，providing a reliable approach for intelligent data mining and a reference for knowledge graph construction in specialized domains.